[tannevaled]

First step on #104: two allocation sources that every blocked factorization hits once per panel.

• GEMM on one worker (internal/kernels/gemm.go). With GOMAXPROCS = 1 (or m = 1), matMul now takes the serial packGemmRows path. It no longer builds the block list, the progress counters and the worker closure, which cost four allocations per call. That path is the one MatMul and the small-product branch already use, and the results are unchanged: the kernels, linalg and root suites pass.
• fastGemm's temporary (linalg/blas.go). The m·n float64 buffer comes from a pool (linalg/scratch.go) and is zeroed on reuse. Buffers above 8 MiB are not kept, so one huge product does not pin its temporary.
• The 9.5 MB linalg/linalg.test binary committed with feat(linalg): numpy.linalg in pure Go #46 is deleted, and *.test is ignored.

Allocation counts, 64×64, from go test -run TestPerf -perf.record with GOMAXPROCS=1 (exact under any load; docs/perf.json and perf-table.md are updated):

row	before	after
linalg.Lstsq / LstsqRcond	276	250
linalg.SVD	265	241
linalg.Pinv / PinvRcond	255	230
linalg.Eigh	166	154
linalg.QR	93	81
linalg.Eig	89	80
linalg.Solve	40	37
linalg.Inv	30	27
linalg.Slogdet	29	28
linalg.Det	25	24
linalg.Cholesky	20	19
Array.MatMul / Array.Dot 128²	13	8

These counts are for one worker, which is how the harness documents its record mode (GOMAXPROCS=1 go test -run TestPerf -perf.record); perf.json's "16 CPUs" is the machine. With several workers the parallel path is unchanged and still costs its four allocations per GEMM call.

TestGemmOneWorkerSameBits pins the new branch. On random (non-integer) data above GemmThreshold, in float64 and float32, it requires GOMAXPROCS=1 to give the same bits as GOMAXPROCS=4, including m = 1. Coverage of the branch therefore no longer depends on stress_test.go's random GOMAXPROCS.

Timing is deferred: the load here is above 8.

What is left, from a memory profile of SVD, QR, Eigh and Eig at 64²:

• The divide-and-conquer merges allocate each of their buffers per merge: bdMerge (linalg/bdc.go, about 12 buffers) and dcMerge (linalg/dc.go, about 7). Removing them needs one workspace passed down the recursion, as LAPACK's dbdsdc/dstedc do.
• newBlockReflector and applyLeft allocate per panel.
• wrap copies every result through ndarray.FromSlice. A zero-copy constructor would avoid it (Refs Wrap: an array over a caller's buffer with given strides (zero-copy, for Julia column-major arrays) #76).

Refs #104

🤖 Generated with Claude Code